Reza2kn/Hes-Shenaas-v0.1

Hes Shenaas v0.1

Hes Shenaas v0.1 is a single-model Persian emotion classifier based on XLM-RoBERTa Large. It predicts one of seven labels:

Label	Meaning
ANGRY	anger
FEAR	fear or anxiety
HAPPY	happiness or joy
HATE	hate or strong aversion
SAD	sadness
SURPRISE	surprise
OTHER	neutral, unclear, or outside the six emotions

Evaluation

The checkpoint was selected using a fixed 1,232-example validation split. The protected 1,151-example test split was evaluated after checkpoint selection.

Split	Examples	Accuracy	Macro F1	Weighted F1
Validation	1,232	73.94%	75.05%	73.76%
Test	1,151	77.58%	76.27%	77.64%

Test F1 by class:

ANGRY	FEAR	HAPPY	HATE	SAD	SURPRISE	OTHER
68.47%	81.03%	86.04%	70.83%	79.77%	72.66%	75.11%

These numbers describe the project's frozen Persian emotion benchmark and should not be assumed to transfer unchanged to other domains.

Usage

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="Reza2kn/Hes-Shenaas-v0.1",
    tokenizer="Reza2kn/Hes-Shenaas-v0.1",
)

print(classifier("امروز واقعاً روز فوق‌العاده‌ای بود"))

For complete class probabilities:

print(classifier("از این وضعیت خیلی عصبانی‌ام", top_k=None))

The model truncates inputs to 128 tokens, matching training and evaluation.

Training

• Base model: FacebookAI/xlm-roberta-large
• Base revision: c23d21b0620b635a76227c604d44e43a9f0ee389
• Training rows: 8,471
• Original cleaned rows: 4,862
• Added public Persian tweets: 3,609
• Added rows were accepted only when two Gemini 3.8 Flash passes agreed and both marked the text self-contained, natural Persian, supported by exact textual evidence, and at least 0.85 confidence.
• Gemini labels were used for training only. Validation and test texts were excluded from API payloads.
• Selection: maximum validation accuracy with earlier epoch retained on ties
• Selected epoch: 5

Limitations

The model is intended for research and product prototyping, especially short informal Persian text. Emotion labels are subjective, and performance can vary on formal writing, code-switching, sarcasm, dialects, long documents, or domains unlike the evaluation data. The OTHER label combines neutral and ambiguous cases. Do not use predictions as the sole basis for high-impact decisions about people.

Integrity

The published model.safetensors SHA-256 is:

43737d1933dd36966f541e9155aa02be1d394ff9f7ff2e475515fc9ed1413030

The metrics and checkpoint were independently rechecked before release.

Licensing and attribution

The Reza2kn fine-tuning modifications are released under the Apache License, Version 2.0. The underlying XLM-RoBERTa Large model remains subject to its MIT license. Both notices apply to this distribution. See LICENSE. Existing third-party copyright, license, and attribution notices remain applicable.

• Upstream: FacebookAI/xlm-roberta-large; declared license: mit.
  • Retained notice: LICENSE-MIT.